Skip to content

Latest commit

 

History

History
273 lines (209 loc) · 11.7 KB

README.md

File metadata and controls

273 lines (209 loc) · 11.7 KB

Maximising tolerance to disturbances via combined control-actuation optimisation for robust humanoid robot walking

Disturbance magnitude (external impulse (N), CoM noise (m)) Baseline robot Gain tuning (GT-PB) Co-design (CD-PB) Co-design (CD-SSDT)
[0, 0] ✅ ✅ ✅
[0, 0] ✅ ✅
[115, 0.015] ❌ ✅
[120, 0.02] ❌ ❌

Quickstart to visualising the simulation result

If you want to jump right into visualising the robots' walking simulation output, you can do that by downloading and opening the file simulation_output.html in your browser.

Installation instructions

Installation using mamba package manager A quick way to install the dependencies is via [mamba package manager](https://mamba.readthedocs.io/en/latest/index.html). To install all the dependencies, 
mamba env create -f environment.yml

To get the correct values corresponding to the hardware of the robot, clone:

git clone https://github.com/robotology/robots-configuration.git@e5a262e

Then clone the current repository:

git clone https://github.com/ami-iit/paper_sathuluri_2024_ral_sso-dt.git

Make sure that the necessary IPOPT solvers are installed and configured correctly. To install the dependencies from IPOPT refer to this or from CasADi via this. The specific HSL packages can also be installed from here.

Usage

Setting up the URDF for the first run

Activate the created environment,

mamba activate walk_release

Before running the actual simulation the urdf file needs to be made Drake compatible by running,

python src/urdf2drake.py
cd src

Now the actual simulation can be run. To visualise the walking robot with pre-defined design variable values run: visualise_robot_simulation.py

Running the point-based optimisation problem

Running a co-design point-based optimisation problem

time python ergoCub_point_based_optimisation.py --generations 500 --population 125 --workers 125 --algo nevergrad_cmaes --seed 42 --add_base_noise 0.02 --add_root_disturbance 3e2 --comment "running point-based optimisation"
Collecting data around the chosen point-optimum

Randomly sampling the space to collect data for training the MLPClassifier

time python collect_data_buffer.py
Training the classifier

Randomly sampling the space to collect data for training the MLPClassifier

time python create_classifier.py
Moving on to the SSDT optimisation

Running the set-based optimisation problem

time python ergoCub_set_based_optimisation.py --generations 500 --workers 120 --population 120 --nas 50000 --nbs 1000 --use_surrogate True --use_compensation True --opt_type ssdt

Computational time per step

Computational time associated

The computational times recorded correspond to an AMD EPYC 9554 64-core Processor with 256 GB RAM running Ubuntu 24.04.1 LTS with a clock speed of 3.76 GHz.

No. Step Details Time
1 Point-based optimisation 500 generations, 120 population ~4 hours
2 Data collection 1e5 samples ~6 hours
3 Classifier training MLPClassifier, 64 neurons, 1 layer ~18 minutes
4 SSDT optimisation 1000 generations, 120 population, 1e4 samples for MCI integration ~1 hour
Convergence of the optimisation for different seeds

Convergence plots for different seeds

You can find the detailed convergence plots for different seeds in this figure.

Ablation studies

Classifier tuning

Hyperparameter tuning of the classifier

Layers Neurons per layer alpha Test set predicted score Test set confusion matrix
1 64 1e-4 0.9554 [[36340, 642], [1071, 347]]
1 64 1e-3 0.9580 [[36489, 493], [1120, 298]]
1 64 1e-2 0.9567 [[36381, 601], [1060, 358]]
1 128 1e-4 0.9472 [[35991, 991], [1038, 380]]
1 128 1e-3 0.9496 [[36050, 932], [1002, 416]]
1 128 1e-2 0.9578 [[36466, 516], [1104, 314]]
1 256 1e-4 0.9418 [[35683, 1299], [936, 482]]
1 256 1e-3 0.9476 [[35974, 1008], [1004, 414]]
1 256 1e-2 0.9585 [[36552, 430], [1165, 253]]
2 64 1e-4 0.9467 [[35987, 995], [1053, 365]]
2 64 1e-3 0.9478 [[36039, 943], [1063, 355]]
2 64 1e-2 0.9510 [[36187, 795], [1087, 331]]
2 128 1e-4 0.9498 [[36078, 904], [1022, 396]]
2 128 1e-3 0.9482 [[36020, 962], [1028, 390]]
2 128 1e-2 0.9519 [[36260, 722], [1126, 292]]
2 256 1e-4 0.9524 [[36260, 722], [1104, 314]]
2 256 1e-3 0.9514 [[36170, 812], [1054, 364]]
2 256 1e-2 0.9429 [[35733, 1249], [943, 475]]
3 64 1e-4 0.9472 [[35999, 983], [1044, 374]]
3 64 1e-3 0.9449 [[35879, 1103], [1014, 404]]
3 64 1e-2 0.9504 [[36135, 847], [1057, 361]]
3 128 1e-4 0.9429 [[35818, 1164], [1030, 388]]
3 128 1e-3 0.9473 [[35999, 983], [1039, 379]]
3 128 1e-2 0.9486 [[36075, 907], [1068, 350]]
3 256 1e-4 0.9549 [[36375, 607], [1123, 295]]
Classifier output vs data

Classifier prediction quality

Sample size (Feasible, infeasible samples) Metric predicted/Ground truth Representative classification
1.28e5 (4838, 123162) 0.245/0.166 Prediction
1.28e4 (4838, 7162) 0.409/0.166 Prediction
1.28e4 (483, 12316) 0.03/0.166 Prediction
1.28e4 (128, 12672) 0.0/0.166 Prediction
1.28e4 (4838, 1279) 0.476/0.166 Prediction
Convergence of the MCI metric vs number of samples used tion

Monte-Carlo integration convergence

Sample size Convergence plot Computation time
1e3 Convergence plot ~5 minutes
5e3 Convergence plot ~15 minutes
1e4 Convergence plot ~28 minutes
5e4 Convergence plot ~96 minutes

Additional Information

Interpreting the magnitude of disturbances

Assuming the robot to be a solid free-floating body, the displacement of the robot under the applied external force can be computed as $\Delta CoM = \frac{1}{2} \frac{F_{applied}}{m_{robot}} t_{applied}^2$. Under the application of the peak force of $300N$, this corresponds to a displacement of $\Delta CoM \approx 0.03m$ which is seen from the plot is around similar orders of the largest error in centre of mass estimation the controller can handle.

Ensuring accuracy of the simulation model
Real1 Real2 Real3
Sim1 Sim3 Sim2

Glossary

Abbreviation Description
CD Co-design
PB Point-based
SSDT Solution space with disturbance tolerance
GT Gain-tuning
DV Design variable
QoI Quantitiy of interest
SSO Solution space optimisation
Symbol Description
$x_K (=x_C) $ Controller gains DVs (joint level and task level)
$\boldsymbol{x}_A$ Actuator DVs (motor torque and gear ratios)
$\boldsymbol{x}_{N}$ Gear ratio DVs
$\boldsymbol{x}_{\tau}$ Motor torque DVs
$\boldsymbol{x}_{SSDT}^*$ Selected design from SSDT

Working with your own walking robot for co-design?

If you want to setup your own robot for control and design co-optimisation using the pipeline described in the paper, have a look at CoMoDO.

Maintainer

This repository is maintained by:
@akhilsathuluri